REVIEW 4 major objections 6 minor
TFCDiff: Robust ECG Denoising via Time-Frequency Complementary Diffusion
T0 review · 4 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash
Pith's one-line read TFCDiff claims that ECG denoising works best when a conditional diffusion model runs in the DCT domain on raw 10-second multi-beat recordings, with a temporal-feature enhancement stream, rather than operating beat-by-beat in the time domain
desk verdict Serious, well-executed DCT-domain diffusion for multi-beat ECG denoising with a genuine external test, but the SimEMG advantage is likely inflated by an untested 50-Hz band limit. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central machinery is conditional diffusion in the DCT domain. A 10-second ECG is DCT-transformed, truncated to the first ~1000 coefficients (below 50 Hz), and both clean target and noisy conditioning input are scaled by a percentile-derived bound (η≈3) from the DC component. The noise predictor is a 1-D U-Net with a Temporal Feature Enhancement Mechanism: Temporal Feature Extraction converts frequency maps to time via zero-padding and IDCT, processes them, and returns via DCT; Temporal Feature Fusion applies time-domain attention at the middle encoder layer. A detour-resample strategy converts maps to time before down/upsampling to avoid aliasing, and the SNR schedule is rescaled by fact
What would settle it
Take raw multi-beat ECG with a simultaneously recorded reference clean lead (e.g., shoulder electrodes), apply TFCDiff without any preprocessing, and measure residual error with and without applying the paper's own bandpass/median/piecewise-baseline pipeline to the reference. If the model's advantage over the strongest non-diffusion baseline shrinks or reverses when the reference is not pipeline-processed, the claimed gain is at least partly an artifact of learning the preprocessing. Additionally, injecting a noise type absent from training (e.g., 50-Hz powerline interference) and observing wh
Extended reading notes
Core claim
The paper's finding is that a conditional diffusion model trained and sampled entirely in the truncated DCT domain—conditioned on the noisy signal's DCT coefficients, with a U-Net that intermittently converts feature maps to the time domain—outperforms time-domain diffusion, autoencoders, and filters on raw 10-second ECG segments. The authors report best results with an SNR-scaling factor of 150, a DC-derived scaling bound, and k-generation averaging. On the synthesized test set, TFCDiff-10 achieves SSD 21.5 au² and ImSNR 14.1 dB, surpassing all eight benchmarks; on an unseen real EMG-contaminated database, even one-generation TFCDiff beats the strongest non-diffusion baseline on all five me
Load-bearing premise
The load-bearing premise is that the filtered, baseline-corrected reference ECG recordings used as ground truth are faithful clean signals, and that amplitude-normalized mixtures of three archived noise excerpts at intensity λ∈[0.2,2] span real ambulatory noise.
Editorial extensions
If this is right
- A single model can denoise raw 10-second multi-beat ECG without beat segmentation or padding, removing a barrier to wearable deployment.
- Training under flexible random mixed noise (random weights for baseline wander, muscle artifact, and electrode motion, intensity λ∈[0.2,2]) yields robustness across noise levels up to the hardest 1.5–2.0 intensity bin.
- With one generation, TFCDiff already outperforms the strongest non-diffusion baseline on an unseen real EMG-contaminated database, suggesting cross-dataset generalization rather than overfitting to synthesized noise.
- Averaging more generations (3, 5, 10) monotonically improves all metrics, so users can trade compute (0.17–1.63 s per 10-s clip) for fidelity.
- Denoised waveforms visually preserve T-wave morphology that time-domain diffusion smooths away, which matters for clinical interpretation of repolarization abnormalities.
Reading between the lines
- If the DCT-domain advantage generalizes, the same truncated-DCT conditioning plus time-frequency fusion could apply to other quasi-periodic biosignals (PPG, EEG, EMG) where morphology preservation is diagnostically valuable.
- The claimed generalization to real EMG noise is tested only against EMG-type interference; testing against real electrode-motion or baseline-wander recordings would clarify whether the model's robustness extends beyond the noise type present in the unseen test set.
- The ground-truth preprocessing dependency suggests a discriminating experiment: compare TFCDiff on raw multi-beat signals against a simultaneously recorded clean lead without the paper's filtering/baseline pipeline; if the margin narrows, the model may be learning the preprocessing rather than true ECG structure.
- The authors note unsatisfactory results for intense EMG noise and suggest flow matching; a direct measure of residual noise power inside the QRS band as a function of input SNR would quantify this ceiling and guide whether a different generative formulation is needed.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes TFCDiff, a conditional diffusion model for denoising 10-second ambulatory ECG signals. The method operates in the DCT domain: clean and noisy signals are transformed by DCT, truncated to the first 1000 coefficients (≈50 Hz at fs=360 Hz), and used as diffusion training targets and conditioning inputs. The noise predictor is a 1D U-Net augmented with a Temporal Feature Enhancement Mechanism (TFEM) that interleaves time-domain and frequency-domain feature streams. Training data are QT Database segments preprocessed with bandpass/median filtering and piecewise-linear baseline removal, corrupted by flexible random mixtures of BW, MA, and EM noise from the NST Database. The model is evaluated on a synthesized QT/NST test set and on the external SimEMG database, with comparisons against eight baselines. Table III reports that TFCDiff-10 achieves the best mean values on all five metrics for the synthesized set, while Table IX shows TFCDiff-3 and TFCDiff-1 outperforming all benchmarks on the SimEMG set. The paper also includes ablations for the DCT-domain design and TFEM, plus generation-time measurements.
Significance. If the results hold, TFCDiff addresses a real gap: multi-beat, raw-sequence ECG denoising without beat segmentation, which is more clinically relevant than single-beat evaluation. The external SimEMG validation is a genuine strength, as is the public code release and the systematic ablation of the time-frequency design. The DCT-domain diffusion idea with TFEM is technically interesting and the reported computational costs are plausible. However, the central SOTA and generalization claims currently rest on a comparison that may be confounded by the hard 50-Hz band-limit inherent to the DCT truncation, and the abstract overstates the synthesized-set result by omitting the ensemble-size dependence. The underlying method is defensible and the issues are addressable with additional control experiments, so the contribution is potentially significant for wearable ECG denoising.
major comments (4)
- [§III-A, §V-B, Table IX] The DCT truncation to 50 Hz gives TFCDiff a structural advantage that is not controlled for. All inputs and outputs are hard band-limited by zeroing DCT coefficients above [50/Δf] (≈1000 of 3600 coefficients at fs=360 Hz). In contrast, all eight baselines operate on full-band time-domain signals. On SimEMG, which contains broadband real EMG noise, simply truncating the DCT of the noisy input and applying IDCT may remove a large fraction of the noise without any learned denoiser. The paper does not include a truncation-only baseline, nor does it band-limit the baselines. Consequently, the claim in Table IX that TFCDiff-3 outperforms all benchmarks on all metrics is confounded: the improvement could be largely due to the fixed 50-Hz cutoff rather than the learned conditional denoiser or TFEM. Please add a control that applies the same 50-Hz truncation to the noisy input (and optionally to
- [Abstract, §V-A, Table III] The state-of-the-art claim in the abstract ('achieves state-of-the-art performance across five evaluation metrics') is unqualified. Table III shows that TFCDiff-1 is worse than TCDAE on SSD, PRD, and CosSim; TFCDiff-3 lags on SSD, PRD, and CosSim; and only TFCDiff-10 leads on all five metrics. The ensemble size k is a critical hyperparameter that determines the headline result. Please qualify the SOTA claim with the ensemble size, and either state TFCDiff-10 explicitly or report the trade-off between k and performance. The conclusion also repeats the unqualified claim.
- [§IV-B] The ground-truth x0 used for training and for evaluating the synthesized set is not the raw QT recording but a surrogate: bandpass/median-filtered, then piecewise-linear baseline-removed within QRS-annotated segments, with Hermite interpolation at junctions. The paper does not validate this surrogate against the original signal or an independent reference. If residual baseline wander, filter ringing, or interpolation artifacts remain, the model is trained to reproduce those artifacts and the reported SSD/MAD/PRD/CosSim/ImSNR measure reconstruction to the processed surrogate, not to the true clean ECG. Please provide a quantitative or visual validation of the preprocessing (e.g., comparison of the surrogate with the raw signal in low-noise segments, or an expert review), and discuss how any residual artifacts affect the interpretation of the intra-dataset results.
- [§VI-A, Table X] The ablation in Table X does not isolate the effect of the 50-Hz truncation from the effect of DCT-domain modeling. The standalone DCT-diffusion row (SSD 146.8, CosSim 0.727) is much worse than the time-domain baseline (SSD 59.5, CosSim 0.889), and adding TFEM reverses this. This suggests that TFEM is compensating for information lost by truncation or by the skewed DCT-coefficient scaling, but the reader cannot tell how much of the final advantage comes from the hard band-limit versus from the learned denoiser. Please include a control where a time-domain model is trained on the same truncated-DCT-reconstructed signals (i.e., band-limited inputs without the DCT diffusion objective), or equivalently, where the time-domain baselines receive a 50-Hz low-passed version of the input.
minor comments (6)
- [§III-A] Duplicate sentence: 'Both transforms must be orthogonal to ensure invertibility.' appears twice consecutively.
- [Throughout] The name 'DesCod' is inconsistently written as 'DesCod' and 'DeScoD' (e.g., Section IV-D vs. Section V-A and reference [18]). Please unify.
- [Table III] CBAM-DAE reports PRD = 1910.425±877.430%, which is physically implausible for ECG denoising; consider explaining this outlier or omitting it from the table with a note, as is already done for later tables.
- [§IV-D] The text says all methods use input length 3600, with exceptions for FCN-DAE and CBAM-DAE at 3584. Please clarify how the outputs are aligned to the 3600-sample evaluation window (padding? cropping?) so that the metric comparison is unambiguous.
- [Fig. 4] The caption of Fig. 4(b) does not identify which curve corresponds to which SNR scaling factor c. Please add a legend.
- [§IV-C] The PRD definition in Eq. (19) uses the denoised signal's mean in the denominator; this differs from some prior definitions that normalize by the clean signal's mean. Please ensure this is intentional and state it clearly, since PRD values are otherwise hard to compare across papers.
Circularity Check
No significant circularity: TFCDiff's claims are empirical benchmark results, externally anchored by held-out test sets and the SimEMG database.
full rationale
The paper's central claims are empirical benchmark measurements, not derivations that reduce to their own inputs. TFCDiff is trained on 33,294 synthesized QT/NST pairs and evaluated on a held-out 4,296-pair test set as well as the external SimEMG database never used in training (Section IV-B, Tables III and IX). No reported metric is a fitted parameter of the model, and the k-generation averaging is an evaluation protocol rather than a prediction. The DCT truncation to 50 Hz and the scaling bound eta from Eq. 9 are transparent preprocessing/hyperparameter choices, with truncation fidelity independently checked by SSD reconstruction error. The choice of c=150 is described as an empirical search ('We empirically find that the diffusion model yields the best performance when c is set to 150'), not as a derived result. Citations such as [30] for DCT-space diffusion and [33] for cross-domain feature fusion provide architectural inspiration but are not invoked as an unverified uniqueness theorem that forces the conclusion. The skeptic's band-limiting concern is a fairness/threat-to-validity issue for baseline comparison, not circularity, because the SimEMG generalization claim is externally grounded. The paper's own stated limitations (Section VI-D) acknowledge gaps such as poor performance under intensive EMG noise and lack of device validation, which further supports that no hidden circular step is being asserted as a derivation.
Assumptions & free parameters
free parameters (7)
- SNR scaling factor c =
150
- DC scaling percentile tau =
1.75 (yields eta ~= 3)
- DCT truncation length =
1000 coefficients (~50 Hz cutoff)
- noise intensity factor lambda =
uniform in [0.2, 2]
- noise mixture weights r,m,n =
r+m+n=1 (sampling distribution unspecified)
- generation ensemble count k =
10 for the headline SOTA (tables also report k=1,3,5)
- diffusion schedule endpoints =
T=50, beta_1=1e-4, beta_T=0.5
assumptions (6)
- standard math DDPM forward process, reverse process, and ELBO-derived L1 training objective (Eqs. 3-8)
- domain assumption Truncated DCT coefficients of ECG are adequately modeled as diffusible data corrupted by unit Gaussian noise
- domain assumption ECG diagnostic information lies below 50 Hz; coefficients beyond index 1000 are negligible
- domain assumption QT recordings after bandpass/median filtering and piecewise-linear baseline removal within QRS-annotated segments are clean ground truth
- domain assumption Amplitude-normalized linear mixtures of NST BW/MA/EM excerpts (Eq. 16) represent real ambulatory noise
- domain assumption Scaling bound eta computed from the DC component alone suffices to normalize the AC coefficients
Cite this review
Pith. "Pith review of TFCDiff: Robust ECG Denoising via Time-Frequency Complementary Diffusion." pith.science (2026). https://pith.science/paper/G2GQP33E
@misc{pith2026251116627,
author = {Pith},
title = {Pith review of: TFCDiff: Robust ECG Denoising via Time-Frequency Complementary Diffusion},
year = {2026},
howpublished = {\url{https://pith.science/paper/G2GQP33E}},
note = {Machine review of arXiv:2511.16627}
}
read the original abstract
Ambulatory electrocardiogram (ECG) readings are prone to mixed noise from physical activities, including baseline wander (BW), muscle artifact (MA), and electrode motion artifact (EM). Developing a method to remove such complex noise and reconstruct high-fidelity signals is clinically valuable for diagnostic accuracy. However, denoising of multi-beat ECG segments remains understudied and poses technical challenges. To address this, we propose Time-Frequency Complementary Diffusion (TFCDiff), a novel approach that operates in the Discrete Cosine Transform (DCT) domain and uses the DCT coefficients of noisy signals as conditioning input. To refine waveform details, we incorporate Temporal Feature Enhancement Mechanism (TFEM) to reinforce temporal representations and preserve key physiological information. Comparative experiments on a synthesized dataset demonstrate that TFCDiff achieves state-of-the-art performance across five evaluation metrics. Furthermore, TFCDiff shows superior generalization on the unseen SimEMG Database, outperforming all benchmark models. Notably, TFCDiff processes raw 10-second sequences and maintains robustness under flexible random mixed noise (fRMN), enabling plug-and-play deployment in wearable ECG monitors for high-motion scenarios. Source code is available at https://github.com/Miroircivil/TFCDiff.
Figures
Figures from the paper (6 more)
Reviewed August 3, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.